Finding the Sweet Spot: Preference Data Construction for Scaling Preference Optimization